JA
J. Agterdenbos
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
2 records found
1
Bridging the Spatio-Temporal Gap in the Trajectory Based Aircraft Landing Problem
From Path Planner to Arrival Manager with Goal-Conditioned Reinforcement Learning
ICAO projects a near-tripling of passenger volumes to 12.4 billion by 2050, yet the resulting bottleneck sits not in the sky but at the runway. Reinforcement learning agents can cut mid-air separation violations by over $99\%$, but treat the airport as a single infinite-capacity sink: traffic concentrates on one runway while others idle, a failure mode termed Runway Overload. This work closes that gap with a Goal-Conditioned Reinforcement Learning hierarchical framework coupling spatial and temporal arrival management. A high-level Manager uses Constrained Position Shifting to allocate runways and assign Required Times of Arrival. A low-level 4D-Worker, trained with Soft Actor-Critic, treats time as a state variable and learns path-stretching manoeuvres to meet arrival times without predefined templates. An Extra Trees regressor bridges the layers with fast arrival-time estimates, avoiding full trajectory rollouts. Evaluated in BlueSky-gym across two training phases and a coordination-evaluation phase, the selected policy reaches success rates above 99.8\%, on-time rates near 99.9\% and tracking error below 0.2 minutes. A heading-augmented observation, intended to preserve the Markov property during holding patterns, instead causes catastrophic failure on closely spaced parallel runways when trained without hindsight relabelling, possibly reflecting a spatial-reachability constraint; the selected policy omits both heading augmentation and hindsight experience relabelling without performance loss. Coupling the Manager with this policy no longer reproduces the runway-overload signature of prior single-sink models under either runway-assignment mode: static assignment splits traffic evenly by construction, whereas dynamic assignment leaves a small residual load imbalance and a separation-compliance cost that widens sharply at low $k$ before plateauing, without a measured benefit in this uniform-demand setting. Across 600,000 simulated aircraft, stalling is effectively absent and no delay propagation between arrivals is observed; under homogeneous-fleet, single-airport conditions the framework therefore addresses the spatio-temporal integration gap in trajectory-based arrival management.
...
ICAO projects a near-tripling of passenger volumes to 12.4 billion by 2050, yet the resulting bottleneck sits not in the sky but at the runway. Reinforcement learning agents can cut mid-air separation violations by over $99\%$, but treat the airport as a single infinite-capacity sink: traffic concentrates on one runway while others idle, a failure mode termed Runway Overload. This work closes that gap with a Goal-Conditioned Reinforcement Learning hierarchical framework coupling spatial and temporal arrival management. A high-level Manager uses Constrained Position Shifting to allocate runways and assign Required Times of Arrival. A low-level 4D-Worker, trained with Soft Actor-Critic, treats time as a state variable and learns path-stretching manoeuvres to meet arrival times without predefined templates. An Extra Trees regressor bridges the layers with fast arrival-time estimates, avoiding full trajectory rollouts. Evaluated in BlueSky-gym across two training phases and a coordination-evaluation phase, the selected policy reaches success rates above 99.8\%, on-time rates near 99.9\% and tracking error below 0.2 minutes. A heading-augmented observation, intended to preserve the Markov property during holding patterns, instead causes catastrophic failure on closely spaced parallel runways when trained without hindsight relabelling, possibly reflecting a spatial-reachability constraint; the selected policy omits both heading augmentation and hindsight experience relabelling without performance loss. Coupling the Manager with this policy no longer reproduces the runway-overload signature of prior single-sink models under either runway-assignment mode: static assignment splits traffic evenly by construction, whereas dynamic assignment leaves a small residual load imbalance and a separation-compliance cost that widens sharply at low $k$ before plateauing, without a measured benefit in this uniform-demand setting. Across 600,000 simulated aircraft, stalling is effectively absent and no delay propagation between arrivals is observed; under homogeneous-fleet, single-airport conditions the framework therefore addresses the spatio-temporal integration gap in trajectory-based arrival management.
Bachelor thesis
(2024)
-
E.G.M. Abbenhuis, J. Agterdenbos, V. Bonanno, M. Camporeale, K. Do Cao, B.A. Kleipool, V.S. Lubbers, A.K. Mielczarek, T. Nakagawa, A.C. Petrache, R.N.H.W. van Gent, P.L.N. Ngo, S. Anand
The report details the design of a wildfire management system consisting of a swarm of electrically-powered unmanned aerial vehicles (UAVs) engaging in pre-, active- and post-fire operations. Specifically, the system monitors key parameters in wildfire detection and spread modelling, simultaneously providing a mobile communication network supporting emergency responders acting on the ground. At an altitude of 850m, necessary measurements can be taken with a three-meter ground resolution. The vehicles have a fixed-wing, single-propeller configuration, allowing a five-hour endurance. The swarm comprises twenty surveillance and five relay units, with five backup units stationed on the ground for contingency management. All necessary hardware for a swarm fits neatly into two standard twenty-foot shipping containers. After taking off from an electronically-activated deployable launch rail, the UAVs are capable of autonomous flight. The swarm intelligence is based on zig-zag flight patterns inside Voronoi sectors generated from wildfire risk maps dynamically updated by the UAVs' sensors. Ground operations are limited to monitoring the system state, as autonomous contingent behaviour is also accounted for. Assuming a maximum fire spread rate of 2.5 m/s, the system is capable of detection in 16 minutes on average. This is a competitive performance compared to existing fire detection methods such as satellites and aeroplanes. After flying, the UAVs return to the ground station and land autonomously using an arresting gear, removing the need for runways and landing gear. A financial analysis of the system reveals a cost of 750,000 USD per swarm, proving its competitiveness in price and performance compared to available market options. It is concluded that such a system is technically and financially feasible and can realistically impart a positive change to wildfire management efforts worldwide.
...
The report details the design of a wildfire management system consisting of a swarm of electrically-powered unmanned aerial vehicles (UAVs) engaging in pre-, active- and post-fire operations. Specifically, the system monitors key parameters in wildfire detection and spread modelling, simultaneously providing a mobile communication network supporting emergency responders acting on the ground. At an altitude of 850m, necessary measurements can be taken with a three-meter ground resolution. The vehicles have a fixed-wing, single-propeller configuration, allowing a five-hour endurance. The swarm comprises twenty surveillance and five relay units, with five backup units stationed on the ground for contingency management. All necessary hardware for a swarm fits neatly into two standard twenty-foot shipping containers. After taking off from an electronically-activated deployable launch rail, the UAVs are capable of autonomous flight. The swarm intelligence is based on zig-zag flight patterns inside Voronoi sectors generated from wildfire risk maps dynamically updated by the UAVs' sensors. Ground operations are limited to monitoring the system state, as autonomous contingent behaviour is also accounted for. Assuming a maximum fire spread rate of 2.5 m/s, the system is capable of detection in 16 minutes on average. This is a competitive performance compared to existing fire detection methods such as satellites and aeroplanes. After flying, the UAVs return to the ground station and land autonomously using an arresting gear, removing the need for runways and landing gear. A financial analysis of the system reveals a cost of 750,000 USD per swarm, proving its competitiveness in price and performance compared to available market options. It is concluded that such a system is technically and financially feasible and can realistically impart a positive change to wildfire management efforts worldwide.